G.M. Behery International Journal of Artificial Intelligence and Expert Systems (IJAE), Volume (4) : Issue (2) : 2013 27 An Automatic Neural Networks System for Classifying Dust, Clouds, Water, and Vegetation from Red Sea Area G.M. Behery behery2911961@yahoo.com Faculty of science, Math.and Comp. Department Damietta University New Damietta, 34517, Egypt. Abstract This paper presents an automatic remotely sensed system that is designed to classify dust, clouds, water and vegetation features from red sea area. Thus provides the system to make the test and classification process without retraining again. This system can rebuild the architecture of the neural network (NN) according to a linear combination among the number of epochs, the number of neurons, training functions, activation functions, and the number of hidden layers. Theproposed system is trained on the features of the provided images using 13 training functions, and is designed to find the best networks that has the ability to have the best classification on data is not included in the training data.This system shows an excellent classification of test data that is collected from the training data. The performances of the best three training functionsare%99.82, %99.64 and %99.28for test data that is not included in the training data.Although, the proposed system was trained on data selected only from one image, this system shows correctly classification of the features in the all images. The designed system can be carried out on remotely sensed images for classifying other features.This system was applied on several sub-images to classify the specified features. The correct performance of classifying the features from the sub-images was calculated by applying the proposed system on some small sections that were selected from contiguous areas contained the features. Keywords: NNs , Image Processing, Classification, Dust, Clouds, Water, Vegetation. 1. INTRODUCTION Remote sensing images provide a general reflection of the spatial characteristics for ground objects. Extraction of land-cover map information from multispectral or hyperspectral remotely sensed images is one of the important tasks of remote sensing technology [1-3]. Precise information about the landuse and land cover changes of the Earth’s surface is extremely important for any kind of sustainable development program [4, 5].In order to automatically generate such landuse map from remotely sensed images, various pattern recognition techniques like classification and clustering can be adopted [6, 7]. These images are used in many applications e.g. for detecting the change in ground cover [8-10], extraction of forest [11-13], and many others [14-16]. NN algorithms are widely used for classifying features from remotely sensed images [17, 18].NN offers a number of advantages over conventional statistical classifiers such as the maximum likelihood classifier. Perhaps the most important characteristic of NN is that there is no underlying assumption about the distribution of data. Furthermore, it is easy to use data from different sources in the NN classification procedure to improve the accuracy of the classification. NN algorithms have some handicaps related in particular to the long training time requirement and finding the most efficient network structure. Large networks take a long time to learn the data whilst small networks may become trapped into a local minimum and may not learn from the training data. The structure of the network has a direct effect on training time and classification accuracy. The NN architecture which gives the best result for a particular problem can only be determined experimentally. Unfortunately, there is currently no available direct method developed